dgl torch incompatability error: Cannot load Graphbolt C++ library

Question

I'm attempting to install and setup dgl in google colab via the DGL install instructions:

!pip install torch
!pip install  dgl -f https://data.dgl.ai/wheels/torch-2.3/cu121/repo.html

Which installs pytorch2.3.0+cu121 and dgl2.3.0+cu121.

However, when I try to import dgl, I run into the following error:

ImportError: Cannot load Graphbolt C++ library

I've tried some other versions of dgl as well, but always run into some sort of import error.

!nvcc --version shows cuda version 12.2

Is the issue a mismatch of the cuda version (12.1 v 12.2)?

Both this thread and this one seem to have run into similar issues.

If anyone can get this working in colab, please share the workbook link.

Answer 1

The import works when I don't install the GPU-specific version of DGL

!pip install dgl -f https://data.dgl.ai/wheels/torch-2.3/repo.html

Not a real fix if you need CUDA, but it implies that the CUDA versioning mismatch between Colab CUDA (12.2) and DGL (looks like it only supports up to 12.1) is the issue